Multimodal image registration using IECC as the similarity measure

Abstract

The registration of images from positron emission tomography (PET) to those from magnetic resonance imaging (MRI) using mutual information is usually effective, but fails occasionally because of small region of overlap, low-activity defects in the PET image, difference in spatial resolution, etc. In this article, the authors propose the pixel-based individual entropy correlation coefficient (IECC) as a new, more accurate and more robust registration criterion. The authors compare it to the current criteria: Mutual information (MI), normalized mutual information (NMI), and the entropy correlation coefficient (ECC). The anatomical region to be registered was the head. A rigid-body registration was used; no deformation was employed. The authors established the effectiveness of IECC by both simulated data and clinical studies using brain fluorodeoxyglucose (FDG) PET and MRI. Both a normal-activity FDG model and a FDG model with a perfusion defect were used for the PET image. Reconstruction by both filtered backprojection and by ordered subset-expectation maximization was investigated. The mean errors and SDs of IECC were 1.17 +/- 0.85 mm for translation and 1.04 +/- 1.28 degrees for rotation in clinical PET. Those of MI, NMI, and ECC were 1.86 +/- 1.22, 1.86 +/- 0.96, and 1.68 +/- 4 1.05 mm for translations and 1.52 +/- 1.84 degrees, 1.74 +/- 1.68 degrees, and 1.70 +/- 1.90 degrees for rotations. The mean errors and SDs of IECC were 1.56 +/- 0.58 mm for translation and 1.46 +/- 1.66 degrees for rotation in clinical PET model with a 30% perfusion defect. Those of MI, NMI, and ECC were 2.81 +/- 1.41, 2.98 +/- 1.80, and 3.29 +/- 2.08 mm for translations and 3.34 +/- 3.800, 2.87 +/- 3.25 degrees, and 3.04 +/- 3.44 degrees for rotations. The mean errors and SDs of IECC were 1.79 +/- 1.04 mm for translation and 1.64 +/- 1.62 degrees for rotation in clinical PET model with a 50% perfusion defect. Those of MI, NMI, and ECC were 3.49 +/- 1.92, 3.57 +/- 2.22, and 3.49 +/- 1.89 mm for translations and 4.12 +/- 4.24 degrees, 3.62 +/- 4.87 degrees, and 3.44 +/- 3.80 degrees for rotations. The significant differences between IECC and each of the other three criteria were found for displacement misregistration in almost all parameters (p < 0.01). Accuracy of the IECC criterion was higher than that of the other criteria, usually in a statistically significant way.